|
| 1 | +// origin: FreeBSD /usr/src/lib/msun/src/k_tan.c */ |
| 2 | +// |
| 3 | +// ==================================================== |
| 4 | +// Copyright 2004 Sun Microsystems, Inc. All Rights Reserved. |
| 5 | +// |
| 6 | +// Permission to use, copy, modify, and distribute this |
| 7 | +// software is freely granted, provided that this notice |
| 8 | +// is preserved. |
| 9 | +// ==================================================== |
| 10 | + |
| 11 | +// kernel tan function on ~[-pi/4, pi/4] (except on -0), pi/4 ~ 0.7854 |
| 12 | +// Input x is assumed to be bounded by ~pi/4 in magnitude. |
| 13 | +// Input y is the tail of x. |
| 14 | +// Input odd indicates whether tan (if odd = 0) or -1/tan (if odd = 1) is returned. |
| 15 | +// |
| 16 | +// Algorithm |
| 17 | +// 1. Since tan(-x) = -tan(x), we need only to consider positive x. |
| 18 | +// 2. Callers must return tan(-0) = -0 without calling here since our |
| 19 | +// odd polynomial is not evaluated in a way that preserves -0. |
| 20 | +// Callers may do the optimization tan(x) ~ x for tiny x. |
| 21 | +// 3. tan(x) is approximated by a odd polynomial of degree 27 on |
| 22 | +// [0,0.67434] |
| 23 | +// 3 27 |
| 24 | +// tan(x) ~ x + T1*x + ... + T13*x |
| 25 | +// where |
| 26 | +// |
| 27 | +// |tan(x) 2 4 26 | -59.2 |
| 28 | +// |----- - (1+T1*x +T2*x +.... +T13*x )| <= 2 |
| 29 | +// | x | |
| 30 | +// |
| 31 | +// Note: tan(x+y) = tan(x) + tan'(x)*y |
| 32 | +// ~ tan(x) + (1+x*x)*y |
| 33 | +// Therefore, for better accuracy in computing tan(x+y), let |
| 34 | +// 3 2 2 2 2 |
| 35 | +// r = x *(T2+x *(T3+x *(...+x *(T12+x *T13)))) |
| 36 | +// then |
| 37 | +// 3 2 |
| 38 | +// tan(x+y) = x + (T1*x + (x *(r+y)+y)) |
| 39 | +// |
| 40 | +// 4. For x in [0.67434,pi/4], let y = pi/4 - x, then |
| 41 | +// tan(x) = tan(pi/4-y) = (1-tan(y))/(1+tan(y)) |
| 42 | +// = 1 - 2*(tan(y) - (tan(y)^2)/(1+tan(y))) |
| 43 | +static T: [f64; 13] = [ |
| 44 | + 3.33333333333334091986e-01, /* 3FD55555, 55555563 */ |
| 45 | + 1.33333333333201242699e-01, /* 3FC11111, 1110FE7A */ |
| 46 | + 5.39682539762260521377e-02, /* 3FABA1BA, 1BB341FE */ |
| 47 | + 2.18694882948595424599e-02, /* 3F9664F4, 8406D637 */ |
| 48 | + 8.86323982359930005737e-03, /* 3F8226E3, E96E8493 */ |
| 49 | + 3.59207910759131235356e-03, /* 3F6D6D22, C9560328 */ |
| 50 | + 1.45620945432529025516e-03, /* 3F57DBC8, FEE08315 */ |
| 51 | + 5.88041240820264096874e-04, /* 3F4344D8, F2F26501 */ |
| 52 | + 2.46463134818469906812e-04, /* 3F3026F7, 1A8D1068 */ |
| 53 | + 7.81794442939557092300e-05, /* 3F147E88, A03792A6 */ |
| 54 | + 7.14072491382608190305e-05, /* 3F12B80F, 32F0A7E9 */ |
| 55 | + -1.85586374855275456654e-05, /* BEF375CB, DB605373 */ |
| 56 | + 2.59073051863633712884e-05, /* 3EFB2A70, 74BF7AD4 */ |
| 57 | +]; |
| 58 | +const PIO4: f64 = 7.85398163397448278999e-01; /* 3FE921FB, 54442D18 */ |
| 59 | +const PIO4_LO: f64 = 3.06161699786838301793e-17; /* 3C81A626, 33145C07 */ |
| 60 | + |
| 61 | +pub(crate) fn k_tan(mut x: f64, mut y: f64, odd: i32) -> f64 { |
| 62 | + let hx = (f64::to_bits(x) >> 32) as u32; |
| 63 | + let big = (hx & 0x7fffffff) >= 0x3FE59428; /* |x| >= 0.6744 */ |
| 64 | + if big { |
| 65 | + let sign = hx >> 31; |
| 66 | + if sign != 0 { |
| 67 | + x = -x; |
| 68 | + y = -y; |
| 69 | + } |
| 70 | + x = (PIO4 - x) + (PIO4_LO - y); |
| 71 | + y = 0.0; |
| 72 | + } |
| 73 | + let z = x * x; |
| 74 | + let w = z * z; |
| 75 | + /* |
| 76 | + * Break x^5*(T[1]+x^2*T[2]+...) into |
| 77 | + * x^5(T[1]+x^4*T[3]+...+x^20*T[11]) + |
| 78 | + * x^5(x^2*(T[2]+x^4*T[4]+...+x^22*[T12])) |
| 79 | + */ |
| 80 | + let r = T[1] + w * (T[3] + w * (T[5] + w * (T[7] + w * (T[9] + w * T[11])))); |
| 81 | + let v = z * (T[2] + w * (T[4] + w * (T[6] + w * (T[8] + w * (T[10] + w * T[12]))))); |
| 82 | + let s = z * x; |
| 83 | + let r = y + z * (s * (r + v) + y) + s * T[0]; |
| 84 | + let w = x + r; |
| 85 | + if big { |
| 86 | + let sign = hx >> 31; |
| 87 | + let s = 1.0 - 2.0 * odd as f64; |
| 88 | + let v = s - 2.0 * (x + (r - w * w / (w + s))); |
| 89 | + return if sign != 0 { -v } else { v }; |
| 90 | + } |
| 91 | + if odd == 0 { |
| 92 | + return w; |
| 93 | + } |
| 94 | + /* -1.0/(x+r) has up to 2ulp error, so compute it accurately */ |
| 95 | + let w0 = zero_low_word(w); |
| 96 | + let v = r - (w0 - x); /* w0+v = r+x */ |
| 97 | + let a = -1.0 / w; |
| 98 | + let a0 = zero_low_word(a); |
| 99 | + a0 + a * (1.0 + a0 * w0 + a0 * v) |
| 100 | +} |
| 101 | + |
| 102 | +#[inline] |
| 103 | +fn zero_low_word(x: f64) -> f64 { |
| 104 | + f64::from_bits(f64::to_bits(x) & 0xFFFF_FFFF_0000_0000) |
| 105 | +} |
0 commit comments